Deploying and Managing HPC Clusters with IBM Platform and Intel Xeon Phi Coprocessor
1 like956 views
Three sentences: IBM Platform Computing software can help businesses maximize the value of compute-intensive applications by managing heterogeneous high performance computing (HPC) clusters with Intel Xeon Phi coprocessors. The software provisions and schedules jobs on nodes with Xeon Phi cards, monitors Xeon Phi metrics, and reports on jobs' Xeon Phi usage. The National Center for Atmospheric Research leverages IBM Platform LSF to manage its complex HPC environment including a Xeon Phi cluster for advanced computing needs.
Deploying and Managing HPC Clusters with IBM Platform and Intel Xeon Phi Coprocessor
- 1. Best Practices in Deploying and Managing HPC Clusters with Intel® Xeon Phi™ Louise Westoby WW Marketing Manager, IBM Platform Computing June 18, 2013
- 2. Business Innovation Stressing IT Objective: Gain competitive advantage • Innovate with more complex applications / simulations / analytics • Long processing limits number of iterations in a given time period • Explosion of data improves results but adds complexity • Delays and high cost of adding new applications • Difficult to use systemsEnd Users / Business Objective: Reduce cost while maintaining service • Infrastructure silos to meet peak service level requirements • CapEx and OpEx budget growth constrained • Infrastructure issues – power/cooling, space, etc. • Rise of lower cost resources (x86) and virtualization • Evolving trend toward heterogeneous, multi-core programming models IT Organizations
- 3. VIRTUALIZED VIEWOF COMPUTE,NETWORKAND STORAGERESOURCES Application Businesses need to overcome infrastructure limitations to maximize the value of compute and data-intensive applications Application Examples • Simulation • Analysis • Design • Big data IT constrained • Long wait times • Low utilization • IT Sprawl IBM Platform Computing Software Big Data / Hadoop Simulation and Modeling Analytics Today Future Make lots of computers look like “one” Prioritized matching of supply with demand Benefits • High utilization • Throughput • Performance • Prioritization • Reduced cost Repeated for many applications and groups • Clusters • Grid • HPC Cloud VIRTUALIZED VIEWOF COMPUTE,NETWORKAND STORAGERESOURCES Faster time to results Use fewer resources HPC Cloud / Cluster Mgmt
- 4. Complete range of technical computing management software to maximize high performance applications Workload and Resource Management Data Management Infrastructure Management Platform LSF Family Batch, MPI workloads with process mgmt, monitoring, analytics, user portal, license mgmt Platform HPC Simplified, integrated HPC management software for batch, MPI workloads integrated with systems Platform Symphony Family High throughput, near ‘real time’ parallel compute and Big Data / MapReduce workloads Big Data / Hadoop Simulation / Modeling AnalyticsApplications Heterogeneous Resources Compute Storage Network Virtual, Physical, Desktop, Server, Cloud Platform Cluster Manager Family Provision and manage Single Cluster (Standard) to Dynamic Clouds (Advanced) General Parallel File System (GPFS) High performance, distributed parallel file system
- 5. System X and Platform Computing: better together Reference Ecosystem – Leverage the tight integration between IBM System x, Platform Computing software and Intel technology RHEL MS System X App App App Q Logic InfiniBand Intel Xeon Intel Xeon Phi Intel Intel Cluster Ready IBM Platform Computing
- 6. Leveraging Platform HPC to properly provision and configure Xeon Phi environment Add Intel MPSS packages to the repository Create provisioning template to include MPSS package Provision all nodes with Xeon Phi cards Generate MPSS configuration on nodes with Xeon Phi Create network bridge & configure Xeon Phi network Start mpss service automatically on system boot up 1. Provision nodes and install MPSS 2. Install Intel® Xeon Phi ®compilers and run time software 3. Configure Platform HPC ELIM
- 7. Levering Platform LSF or Platform HPC to simplify scheduling of Intel® Xeon Phi™ jobs • Job can be submitted by specifying the following metrics: – Number of Xeon Phi cards required on each node – Any metrics the Xeon Phi ELIM collects • Job will be placed on nodes with available Xeon Phi cards that meet the resource requirements – Numerate Xeon Phi card on a node allowing multiple jobs running on the same node using designated cards • Agnostic to Xeon Phi execution mode (offload, native, etc.) • Job information – Indication of which Xeon Phi cards are used Collecting Xeon Phi Metrics • Total number of cards per node • Number of cores per accelerator • Core temperature (Celsius) • Frequency (GHz) • Total power (Watts) • Total Free memory (MB)
- 8. Cluster Node Platform HPC monitoring system • Single agent for both resource monitoring and resource management • Based on 20 years of Platform technology – Light weight and small footprint – Scalable – Robust – Extendable – Fully automated failover • Added monitoring metrics shown in Platform HPC web GUI automatically • Added monitoring metrics can be used to define alerts LIM Xeon Phi ELIM GPU ELIM Other ELIMs Management Node Master LIM PERF: Monitoring & Reporting Master Scheduler
- 9. Mudpot: Intel® Xeon Phi™ Cluster is used for advanced computing at the NCAR Wyoming Supercomputing Center 9
- 10. IBM Platform LSF Leveraged at NCAR to manage complex, heterogeneous compute environment • From user POV there is one place to submit jobs, regardless of resource • Different queues depending on job type (e.g. regular, bigmem, gpgpu) • Allows multistage jobs to run on multiple resources – Large model run on Yellowstone – Dependent Data-Analysis Run on Geyser • Sharing between projects managed transparently 10
- 11. Thank you!